AMD Presentation For Linux Kernel Summit
Richard A. Brunner AMD Fellow
July 2005 Progress Report
Linux is moving into the mainstream with AMD64 technology
• AMD64+Linux is penetrating deeper into the Data Center – Demanding mainframe/UNIX functionality: 64-bit, NUMA, multi- core, and virtualization – Requiring solutions to infrastructure issues: more power management, security, and manageability – Requesting innovation without disruption: evolution as opposed to revolution - need to maintain compatibility and stability – Using servers and workstations as the proving ground: Linux must do well in these area before they move to Linux on the desktop • AMD continues the trend of openly providing early technical information on our products to the developer community for feedback.
Page 2 July 2005 Linux Kernel Summit AMD’s Technology Roadmap Technology Roadmap
N=Today (N+1) (N+2) (N+3)
Page 4 July 2005 Linux Kernel Summit 130nm to 90nm Performance & Power
Page 5 July 2005 Linux Kernel Summit 65nm Progress
Page 6 July 2005 Linux Kernel Summit Post-45nm Research Begun and Processing
Page 7 July 2005 Linux Kernel Summit AMD’s Processor Roadmap Introducing AMD64 Dual-Core Processor
• Two AMD Opteron™ CPU cores on a single die, each with 1MB L2 cache • 90nm, ~205 million Core 0 transistors* 1-MB L2 – Approximately same die size as 130nm single-core AMD Opteron processor* • 95 watt power envelope fits Northbridge into 90nm power infrastructure • Retains compatibility with existing 32-bit and 64-bit x86-base software 1-MB L2 • Introduced with “K8” Core 1 Revision E core in April 2005
*Based on current revisions of the design
Page 9 July 2005 Linux Kernel Summit Designed From The Start To Add Second Core
• Shared Northbridge – 3 HyperTransport™ technology links – Dual-channel (128 bit) DDR i/f Existing AMD64 • AMD Opteron™ CPU with Direct Processor Design Connect Architecture was designed as CMP from the start 1MB 1MB – Second port on SRI, request L2 Cache L2 Cache management, two APICs • Two complete CPU cores SRI –SMP model Core 0 Core 1 – Simpler, less-restrictive X-bar programming model than “logical core” approach DDR1 DRAM HyperTransport™ – No need to “pause” one core to give Interface Links 0,1,2 other exclusive use of shared resources
Page 10 July 2005 Linux Kernel Summit AMD Dual-Core Technology
AMD Athlon™ 64 X2 Dual-Core Processor (Announced June 2005) Model # Freq L2 Cache 4800+ 2.4 Ghz 1 MB + 1 MB 4600+ 2.4 Ghz 512KB + 512KB Desktop 4400+ 2.2 Ghz 1 MB + 1 MB 4200+ 2.2 Ghz 512KB + 512KB
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_8796_9240,00.html
AMD Opteron™ Processor Dual-Core Models (Announced on April 21, 2005) Freq 1-way Up to 2-way Up to 8-way 1.8 GHz Model 165 Model 265 Model 865 2.0 GHz Model 170 Model 270 Model 870 Server/Workstation 2.2 GHz Model 175 Model 275 Model 875
http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_9485_13041%5E13076,00.html
Page 11 July 2005 Linux Kernel Summit Dual-Core Performance/Watt
• SPECweb® 99_SSL Secure Web Connections Example. • Data Center rack space and power budgets are often fixed. • Perf/Watt focus maximizes use of resources. • Typical 48U Rack has 9KVA of Power.
Page 12 July 2005 Linux Kernel Summit AMD Direct Connect Architecture + Dual-Core
DDR1 16x16cHT Opteron Opteron MEM 800 800 cHT MEM K8 cHT BW DDR BW REV (MHz) 16x16 (MHz) 1-ch/2-ch 1-w/2-w (GB/s) (GB/s) Opteron Opteron 800 800 CG HT1- 3.2/6.4 DDR1 3.2/6.4 800 -400 16x16 E HT1- 4.0/8.0 DDR1 3.2/6.4 1000 -400 PCI-E CORE 0 F HT1- 4.0/8.0 DDR2 Chipset 1000 PCI-E
South CORE 1 Bridge
Page 13 July 2005 Linux Kernel Summit SSE3 Support
• AMD K8 Revision “E” and • ADDSUB[PD,PS] xmm1, xmm2/m128 newer are designed to – Provides interleaved packed add and support SSE3 subtract • Supports SSE3 • FISTTP m16int/m32int/m64int instructions reported by – Like FISTP but with forced truncation CPUID.SSE3 feature flag • HADD[PD,PS] xmm1, xmm2/m128 – Horizontal Adds • Ten new SSE • HSUB[PD,PS] xmm1, xmm2/m128 instructions and one new – Horizontal Subtracts x87 instruction (13 total • LDDQU xmm, m128 opcodes). – Special 128-bit Unaligned load • Monitor/Mwait planned • MOV[D,HD,LD]DUP xmm1, xmm2/m64 in 2007 – Move and Duplicate some elements • CMPXCHG16B planned in 2006
Page 14 July 2005 Linux Kernel Summit Desktop/Workstation Roadmap
Page 15 July 2005 Linux Kernel Summit Server/Workstation Roadmap
Page 16 July 2005 Linux Kernel Summit Planned 2006 Processor Features
• Multi-Core capable • DDR2 support • RDTSCP – see next slide • CMPXCHG16B – compare 16bytes, exchange 16-bytes • Correctable Machine-Check Exception Thresholding • HW Virtualization support (AMD “Pacifica”)
Page 17 July 2005 Linux Kernel Summit RDTSCP: Read Serialized TSC Pair
• New instruction, similar to RDTSC: – Returns 64-bit TSC value in %edx:%eax – Is a serializing operation -- prevents speculative reads of TSC – Returns TSC_AUX[31:0] MSR in %ecx at same time as TSC ¾ OS initializes TSC_AUX to meaningful value ¾ Atomicity ensures no context switch btw read of TSC & TSC_AUX. – Availability determined by new extended CPUID feature flag • Allows TSC and OS-supplied value (such as CPU number) to be read atomically in a serializing way in user mode. – TSC rates between CPUs in MP-system may vary – Linux can put CPU number in TSC_AUX so user-mode get- time-of-day knows which per-cpu adjustments to use to fix- up TSC value.
Page 18 July 2005 Linux Kernel Summit Planned 2007 Processor Features
• Multi-core capable • DDR3 support • 1-GB pages – see next slide • 48-bit Physical Addressing – see later slide • Greater than 32-socket support • P-state Invariant TSC (APIC Timer is already) • P-state Fire-n-Forget • Monitor/Mwait • Shared L3-cache • Further Virtualization extensions
Page 19 July 2005 Linux Kernel Summit 1 Gigabyte Pages & 48-bit Physical Addresses
64-bit 63 4847 39 38 30 29 0 VA Sign-Extend PML4-O PDP-O Offset
CR3 PML4E PDP
Page Map Page Dir Level 4 Pointer Table Table
Physical 47 30 29 0 Address Page PA Offset
Plan is for Physical Address in PTEs to be 48 bits for all page sizes.
Page 20 July 2005 Linux Kernel Summit Virtualization Discussion AMD Virtualization Directions
• AMD “Pacifica”: HW-Virtualization-Assist. Base features planned launch in 2006 Generation • Primary components of Architecture: – Host/guest management hardware support – Event Injection ¾ Eliminates need for VMM code to emulate x86 exception delivery ¾ Designed to reduce VMM development time significantly – Nested Page Tables ¾ Designed to improve VMM performance, and reduce overhead ¾ Helps reduce VMM complexity
Page 22 July 2005 Linux Kernel Summit Core “Pacifica” Architecture: VMRUN
• Virtualization based on Virtual Machine Run ( VMRUN) instruction • VMRUN executed by host causes the guest to run • Guest runs until it exits back to the host • Host resumes at the instruction following VMRUN Host instruction Stream while (1) { VMCB // Do World Switch Data rAX = &VMCB Guest instruction Stream Struct VMLOAD(rAX)
while (running_VMM) { VMRUN(rAX) switch (exitcode) { // handle intercept // within VMM context } Intercepts VMSAVE(rAX) }
Page 23 July 2005 Linux Kernel Summit Core “Pacifica” Architecture: Intercepts
• Guest runs until: – It performs an action that causes an exit to the host – It explicitly executes the VMMCALL instruction • The VMCB for a guest has settings that determine what actions cause the guest to exit to host – These intercepts can vary from guest to guest – Two kinds of intercepts ¾ Exception & Interrupt Intercepts ¾ Instruction Intercepts – Rich set of intercepts allow the host to set customize each guest’s privileges • Information about the intercepted event is put into the VMCB on exit
Page 24 July 2005 Linux Kernel Summit Nested Paging
• CPU maps each Guest_PA to Host_VA and then translates to Host_PA • CPU builds compound gVA_to_hPA TLB entries (guarded by ASID) • Far more efficient than “Shadow Page Tables”, all handled by CPU
Guest Translation Host Translation gPA = gen_PML4(gCR3,gVA); hPA = hTRANS( hVA = gPA ); entry = MEMORY[ hPA ];
gPA = gen_PDP(gVA, entry); hPA = hTRANS( hVA = gPA ); entry = MEMORY[ hPA ];
gPA = gen_PDE(gVA, entry); hPA = hTRANS( hVA = gPA ); entry = MEMORY[ hPA ];
gPA = gen_PTE(gVA, entry); hPA = hTRANS( hVA = gPA ); entry = MEMORY[ hPA ];
gPA = gen_PA(gVA, entry); hPA = hTRANS( hVA = gPA );
Page 25 July 2005 Linux Kernel Summit Challenges / Issues Multi-core Numbering
• Assume system has non-power-of-two number-of-cores in at least 1 processor due to design or retirement of bad core(s). – How to tell OS? How to keep “sanity” in core/processor bit masks? • BIOS calculates “Rounded Number of Cores” (RNC): – RNC = 2^ceil( log2(Number_of_Cores) ) • BIOS assigns APIC IDs of each processor’s cores to an RNC- aligned block of IDs: – APIC_ID[ proc=i, core=j ] = RNC * (OFFSET + i) + j
• Example: 2-processor system Proc Core APIC ID – proc 0 has 3-cores 0 0 0x8 = 4*(2+0) + 0 – proc 1 has 4-cores 0 1 0x9 = 4*(2+0) + 1 – RNC = 4 on all cores 0 2 0xA = 4*(2+0) + 2 rsvd = 4*(2+0) + 0 Want APIC_ID[M:0] to always specify core 1 0 0xC = 4*(2+1) + 0 Initial 1 1 0xD = 4*(2+1) + 1 APIC ID: pppp … cccc Want APIC_ID[N:M+1] to always specify processor 1 2 0xE = 4*(2+1) + 2 1 3 0xF = 4*(2+1) + 3
Page 27 July 2005 Linux Kernel Summit Multi-core Numbering (cont)
• OS should use same process to discover topology of processors & cores. • OS can not assume that BSP’s CPUID.number_of_cores is same for all processors. • OS can assume that RNC calculated on any processor is same for all processors. • BTW, APIC ID and IOAPIC ID space are disjoint spaces, so the IDs can overlap – There is no logic driven by IOAPIC ID on AMD K8 chipsets
Page 28 July 2005 Linux Kernel Summit Other Challenges
• Mixture of physical and logical cores • UEFI and other BIOS Adventures • PCI Express® Extended Configuration Space • Std Kernel-supported Performance Counter mechanism – don’t steal NMI • 1-GB Page support • Number of threads per-process • HW-assisted Virtualization
Page 29 July 2005 Linux Kernel Summit Trademark attribution and cautionary statement
• AMD, the AMD Arrow logo, AMD Athlon, AMD Opteron, AMD Sempron, AMD PowerNow!, and combinations thereof, and Geode are trademarks of Advanced Micro Devices, Inc. HyperTransport is a trademark of the HyperTransport Technology Consortium. PCI-E is a trademark and PCI Express is a registered trademark of PCI-SIG in the U.S. and/or other jurisdictions. SPEC is a trademark of Standard Performance Evaluation Corporation. Other names used are for informational purposes only and may be trademarks of their respective owners. © 2005 Advanced Micro Devices, Inc. All rights reserved. CAUTIONARY STATEMENT This presentation contains forward-looking statements, which are made pursuant to the safe harbor provisions of the Private Securities Litigation Reform Act of 1995. Forward-looking statements in this presentation relates to, among other things, product and technology introduction schedules; and planned product architecture features. Forward-looking statements involve risks and uncertainties that could cause actual results to differ materially from the company’s current expectations. Risks include the possibility that the company may not achieve its current product and technology introduction schedules; that global business and economic conditions will worsen resulting in lower than currently expected sales; that Intel Corporation’s pricing, marketing programs, product bundling, new product introductions or other activities targeting the company’s processor business will prevent attainment of the company’s current plans; that demand for personal computers and, in turn, demand for the company’s processors, will be lower than currently expected; that adoption of AMD64 products by OEMs will not continue as expected; that the company will not be able to meet demand for its products; that the company will not be able to raise sufficient capital to enable it to establish leading- edge capacity to maintain its market leadership positions; and that solutions providers will not timely provide the infrastructure to support the company’s AMD64 technology.
Because the company’s actual results may differ materially from our plans and expectations today, we encourage you to review the company’s filings with the Securities and Exchange Commission, including but not limited to the risks and uncertainties contained in our Annual Report on Form 10-K for the year ended December 26, 2004, and our Quarterly Report on Form 10-Q for the quarter ended March 27, 2005.
Page 30 July 2005 Linux Kernel Summit